Semantic role labeling for portuguese - a preliminary approach

Abstract

Currently there are increasingly more private and academic publications in the form of digital content on the Internet making extremely difficult to extract and maintain the content information manually. Normally, these tasks follow approximations based on natu- ral language processing. This paper presents a preliminary approach for obtaining a semantic role labeler for Portuguese, a little explored as- pect of natural language processing for this language. The approach was evaluated for the 3 most frequent semantic roles (relation, subject and object) with a subset of Bosque 8.0 corpus. The same approach was ap- plied to an English corpus – the CONLL’2004 one and its results were compared to the ones obtained on the CONLL’2004 shared task. At the same time it presents BosqueUE, a Portuguese corpus for semantic role labeling that can be the basis material for future research in the area. This corpus has the same format as the CONLL’2004 one, facilitating multi-language evaluations.